Firework: Efficient and Transparent DSM with Hardware-Coherent CXL Shared Memory
TL;DR
Firework is introduced, a DSM system co-designed across the runtime and OS kernel that transparently and dynamically manages page placement between local DRAM and hardware-coherent CXL memory, enabling unmodified multi-threaded C/C++ applications to execute across multiple servers.
Abstract
Multi-threading offers the most straightforward path to accelerating applications with extra resources, but extending this capability beyond a single physical server remains difficult— it typically demands significant rewrites of application logic and reliance on specialized systems tailored to particular distributed workloads. This paper introduces Firework, a DSM system co-designed across the runtime and OS kernel that transparently and dynamically manages page placement between local DRAM and hardware-coherent CXL memory, enabling unmodified multi-threaded C/C++ applications to execute across multiple servers. A central challenge lies in supporting both transparency and efficiency. For transparency, Firework maintains a partitioned global address space and coordinates page-level sharing and migration across servers without application-level annotations. For efficiency, Firework dynamically shares pages on demand and unshares them as access patterns change, minimizing the footprint of the coherent shared-memory region. Extensive evaluation across a diverse suite of applications demonstrates up to 10.1× performance gain against a shared-heap baseline and promising scalability results.